But, no local AMD hardware means constant SSH juggling. Upload code, compile remotely, run rocprof, download results, repeat. We were spending more time managing infrastructure than optimizing kernels.

Context: DigitalOcean's AMD GPU droplets ship with ROCm preinstalled. At $1.99/hour for MI300X access, it's cheaper than buying hardware. We're currently competing in the GPU mode kernel optimization competition. We spend a good chunk of our time setting up the infra to profile these kernels.

Solution: Built Chisel to make AMD GPU development feel local. One command spins up a droplet, syncs your code, runs profiling with rocprof, and pulls results back. It handles the SSH, rsync, and teardown automatically.

  Key features:
  - chisel up creates MI300X droplet in seconds
  - chisel sync pushes only changed files
  - chisel profile kernel.cpp compiles, profiles, downloads traces
  - chisel pull grabs artifacts back to local
  - Auto cleanup prevents zombie droplets

Available on PyPI: pip install chisel-cli

Source: https://github.com/anthropics/chisel

Would love feedback, especially from people doing GPU work and exploring AMD alternatives to NVIDIA.

If you're interested in contributing, we'd welcome any help. our general short-term direction is to add grafana, concurrent runs, better error handling, and support for other cloud providers.